SGBCD: A Data Discretization Algorithm Based on Granular-ball Computing and Attribute Significance

As an important method in data preprocessing, discretization can effectively reduce the size of data, generate concise semantic representation, obtain valuable knowledge and information contained in big data, which is of great significance in the fields of data mining and machine learning. Nevertheless, most of the traditional discretization algorithms don't consider the distribution of sample and difficulty in setting parameters during interval partitioning, which leads to a decrease in efficiency and classification accuracy. A data discretization algorithm based on granular-ball computing and attribute significance is proposed in this paper, which is a multi-granularity method. Firstly, by introducing the granular-ball computing, the difficulty of parameter optimization is reduced and the efficiency of the algorithm is improved. At the same time, interval partitioning adaptively fits the original data distribution, reducing the occurrence of uneven interval partitioning and improving classification accuracy. Then, the randomness of attribute selection leads to instability in interval partitioning results. By introducing attribute significance and prioritizing the selection of attributes with high importance, this paper further improves the classification accuracy. Compared to other excellent discretization algorithms in the experiments, the proposed algorithm shows an ideal performance.